Final answers hide the path
A run can look successful after a bad tool choice, an ignored constraint, or a recovery that only worked by chance.
Local-first observability for agent systems
Find what keeps failing.
Turn it into an eval.
Perseval finds recurring failures across agent runs, shows the exact evidence, and turns verified patterns into reviewable regression evals.
Local-firstDeterministic by defaultOpen source
Start with recurring failure groups across runs, then open the evidence behind the pattern.
A plausible answer can conceal a broken run.
Trace viewers isolate incidents
One-run-at-a-time debugging makes it hard to see which failures recur across sessions, builds, and agent versions.
Fixes rarely become evals
You find the bug, ship the fix, and three agent versions later it is back, because nothing ever tested for it.
From noisy traces to a decision you can defend.
One investigation loop, four moves. Each one keeps the evidence attached.
See
Start with recurring failure groups ranked by severity, recurrence, affected runs, trend, and recovery.
Understand
Review the diagnosis, representative examples, and the execution shapes behind the pattern.
Verify
Inspect expected versus observed behavior and jump to the exact evidence spans in the full trace.
Create
Turn one or several confirmed patterns into a draft eval definition with evidence attached.
From raw traces to reviewable evals.
Find what keeps failing
Automatically analyze finalized traces and group recurring failures by exact signature. Rank them by severity, frequency, affected runs, trend, and recovery.
Deterministic by defaultVerify every diagnosis
Inspect expected versus observed behavior, highlighted evidence spans, telemetry gaps, and representative examples before accepting a finding.
Exact evidence and provenanceCreate reviewable evals
Generate draft eval candidates from one or several failure groups, complete with cases, proposed behavior, rubric, grader, and evidence packet.
Human approval requiredCollect without replatforming
Ingest OTLP/HTTP JSON or protobuf, gzip-compressed payloads, or local trace files. Keep project, environment, build, and session identity explicit.
Local-first ingestionExplore complex agent runs
Navigate virtualized runs and lazy trace trees. Search errors and inspect agent roles, events, links, attributes, and bounded payload previews.
Built for large multi-agent tracesCompare where runs diverge
Align baseline and candidate execution, preserve both run identities, and move from their common prefix to the first meaningful difference.
Run-to-run clarityNot yetPerseval creates and reviews eval definitions today. Running the accepted evals is on the roadmap.
Run locally, then start with OTLP or a trace file.
Follow one failure from first trace to accepted eval.
Five stops, one investigation. Project, environment, and build identity carry through every screen, so nothing gets lost between finding a failure and testing for it.
01 / Collect
Start with a named project and a healthy source.
Keep environment, build, and session identity explicit before traces arrive. The Runs view preserves those boundaries when you move between agent versions.
02 / Verify
Read the diagnosis next to the rows that produced it.
Expected behavior, observed behavior, impact, provenance, and the next action stay above the exact rows that established the finding.
03 / Explore
Keep agent topology and timing intact.
Move between a hierarchical tree and the loaded timeline without flattening planner, browser, tool, and verifier roles into one ambiguous list.
04 / Create
Turn confirmed groups into eval definitions you can review.
Build candidates from representative cases, inspect the proposed rubric and grader, and keep evidence provenance visible before accepting anything.
05 / Compare and control
Prove whether behavior actually changed.
Align a baseline and candidate run to see exactly where execution diverged, or confirm it did not. Then manage storage, privacy, and AI settings without leaving the app.
Start with the failure. Keep the trace as evidence.
Conventional trace viewerPerseval
Starting pointChoose one runRanked recurring failures
ScopeOne trace at a timeAcross runs, builds, and sessions
EvidenceManual span searchDiagnosis linked to exact spans
Multi-agent contextOften flattenedRoles and topology preserved
OutcomeInvestigation notesReviewable eval definition
No API key required.
Grouping, ranking, and evidence are deterministic. Perseval finds recurring failures without calling a model. Add embeddings, semantic judging, or cohort labels when they earn their cost, with provider, model, and output provenance always on record.
Deterministic detectorsOptional embeddingsVisible model provenanceLocal provider health
Give coding agents structured evidence.
Query projects, runs, failure groups, findings, evidence, trace context, and eval candidates over MCP without scraping the interface or rebuilding context from screenshots.
Explore the MCP tools →Built like a database, not a dashboard.
SQLite journalDurable control state and ingestion history
DuckDB projectionsFast analytical views over collected traces
Content-addressed payloadsCompressed storage with bounded reveal
Revision-aware tracesDeduplication, corrected spans, and reopened traces
Preserved contextEvents, SpanLinks, and typed attributes
Visible pipeline healthQueue, journal, projection, and analysis status
Debugging your agents shouldn't require an expensive observability platform.
Perseval is free and open source. Your traces live in a SQLite and DuckDB workspace on your own machine. No per-seat pricing, no data leaving your laptop, no procurement meeting. Point your OTLP exporter at localhost and start.
What engineers usually ask first.
Is Perseval another trace viewer?
No. A trace viewer starts from one run; Perseval starts from the failures that recur across runs and ends with an eval you can review. The trace is the evidence, not the product.
Does it require an AI provider?
No. Failure grouping, ranking, and evidence work without any model. Embeddings, semantic judging, and cohort labels are optional extras, and every model-generated output records which provider and model produced it.
Where does my trace data go?
Perseval is local-first. It uses a durable SQLite journal, DuckDB analytical projections, and compressed content-addressed payload storage in your workspace.
Can it handle multi-agent traces?
Yes. It preserves agent roles, hierarchy, events, links, and typed attributes instead of flattening every participant into one ambiguous list.
Does Perseval run the evals it creates?
Not yet. Perseval currently creates and reviews eval definitions. Execution of accepted evals is on the roadmap.